California vetoes the frontier model safety bill and enacts a training data disclosure requirement for 2026

2024-09-29US Policy

Veto Message on Senate Bill 1047, Governor Gavin Newsom to the Members of the California State Senate (Sept. 29, 2024); Assembly Bill 2013 (Irwin), Generative artificial intelligence: training data transparency, Stats. 2024, ch. 817 (chaptered Sept. 28, 2024), codified at California Civil Code section 3111

Governor Newsom returned Senate Bill 1047 without his signature. The bill would have required developers of large artificial intelligence models to put safeguards in place, overseen by a new Board of Frontier Models. The veto message objects that "by focusing only on the most expensive and large-scale models" the bill could give the public a false sense of security, and that "smaller, specialized models may emerge as equally or even more dangerous than the models targeted by SB 1047." Assembly Bill 2013 was chaptered the day before as Chapter 817 of the Statutes of 2024. It requires the developer of a generative artificial intelligence system made publicly available to Californians to post on its website, on or before January 1, 2026 and before each new release or substantial modification, "a high-level summary of the datasets used in the development" of the system: their sources or owners, how they further its purpose, the number and types of data points, whether they include copyrighted, trademarked or patented material or are entirely public domain, whether they were purchased or licensed, whether they include personal or aggregate consumer information, how they were cleaned or modified, when they were collected and first used, and whether synthetic data was used. It reaches systems released on or after January 1, 2022, with exemptions for security and integrity, aircraft operation and federal national security systems.

What it changesFor a patent program, the disclosure that matters is the one that passed. From January 1, 2026, a developer serving California must publish a description of its training datasets, their licensing status, the processing applied to them and its use of synthetic data, and that description is a printed publication available to every examiner and competitor from the day it is posted. Any application whose claims rest on the data pipeline, whether synthetic data generation, filtering and deduplication, licensing-aware sampling or mixing ratios, should be on file before the posting, and because the obligation recurs before each new release or substantial modification, patent filings should be scheduled against the release calendar. Patent counsel should review each posting so that it says what the statute requires without describing the mechanism. Copyright plaintiffs will read the same document, a record of which datasets were licensed and which were not, as closely as examiners will. Fifteen months is enough time to set up that review.

More on AI regulation and policy · Copyright and training data

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.